In dense, “noisy” environments, autonomous flying devices have so far only been able to navigate safely with caution. But a new development has shown that this is no longer a problem. Typically, we like to lull ourselves to sleep by saying that, while it’s great to see artificial intelligence advancing, for example, even the most advanced algorithms can only shamefully take a back seat to the complex cognitive abilities of humans, which offer solutions to a myriad of problems – for now.
Something similar could be said about self-driving drone technology. These devices have become more complex, smarter and more autonomous over the years, but they have been unable to emulate the precision and speed of a professional human pilot guiding his machine over difficult terrain with obstacles. But now a joint team from the University of Zurich and Intel Labs has come up with something that could start to challenge human superiority in this area too.
The demonstration video below shows the amazing speed at which the experimental device, which is in fully autonomous mode, zips through the trees. Watching the footage, one is reminded of the chase scene on the moon of the planet Endor from Return of the Jedi. Knowing the outcome, the Imperial soldiers might have been better off not driving their shuttles themselves, but relying on the Swiss algorithm.
For details of the project, see the publication in the journal Science Robotics. The initial problem was that self-guiding drones are unlikely to work in complex, terrain-challenged and possibly rapidly changing environments because of their sluggishness.
To overcome this, the research team turned to those who had already proven themselves in the field: professional human pilots. The algorithm responsible for manoeuvring the drone was taught exclusively in a simulator, but a professional human teacher was on hand to help in the field. Not only did the neural network learn at lightning speed how a human pilot manoeuvres in a difficult virtual environment, but it was also able to apply the skills it had learned in the real world without having “seen” the environment first.
The researchers are confident that their system, capable of speeds of up to 40 kilometres per hour in tests, could eventually be used in forests, industrial environments and disaster areas, whether it’s searching for missing people or navigating dangerous terrain for humans. At the same time, of course, it could also be used routinely in future military interventions to hunt people – say, in the always challenging urban environment.